Update examples for native SegMask and SegParser (#850)

* Migrate DeepLab examples to native SegmentationMask

* Migrate dino tracking to native SegmentationMask

* Use native visualizer for YOLO-P outputs

* Update generic example segmentation overlay docs

* Update YOLO-P visualizer docs

* Temporarily target DAI 3.8 and depthai-nodes main

* Update generic example instance segmentation docs

* Address PR review follow-ups

* Use released depthai-nodes version

Author: rolandocortez

apps/dino-tracking/backend/requirements.txt

@@ -1,5 +1,5 @@
-depthai==3.7.1
-depthai-nodes==0.5.0
+depthai==3.8.0
+depthai-nodes==0.6.0
 opencv-python-headless~=4.10.0
 python-dotenv
 python-box

apps/dino-tracking/backend/src/annotations/outlines_overlay_node.py

@@ -1,7 +1,6 @@
 import cv2
 import depthai as dai
 import numpy as np
-from depthai_nodes.message import SegmentationMask
 from depthai_nodes.node.base_host_node import BaseHostNode
 
 
@@ -31,12 +30,8 @@ def process(self, frame_msg: dai.ImgFrame, segmentation: dai.Buffer):
             self.out.send(frame_msg)
             return
 
-        assert isinstance(segmentation, SegmentationMask)
-        mask = getattr(segmentation, "mask", None)
-
-        if mask is None:
-            self.out.send(frame_msg)
-            return
+        assert isinstance(segmentation, dai.SegmentationMask)
+        mask = segmentation.getCvMask()
 
         frame = frame_msg.getCvFrame()
         H, W = frame.shape[:2]

apps/dino-tracking/backend/src/object_selection/mask_selection_node.py

@@ -1,6 +1,5 @@
 import depthai as dai
 import numpy as np
-from depthai_nodes.message import SegmentationMask
 from depthai_nodes.node import BaseHostNode
 
 
@@ -31,10 +30,8 @@ def clear_selection(self) -> None:
         self._pending_click = None
         self._selected_mask = None
 
-    def process(self, segmentation: dai.Buffer):
-        assert isinstance(segmentation, SegmentationMask)
-
-        segmentation_mask = segmentation.mask.astype(np.int32)
+    def process(self, segmentation: dai.SegmentationMask):
+        segmentation_mask = segmentation.getCvMask().astype(np.int32)
 
         if self._pending_click:
             segment_id = self._map_click_to_segment(
@@ -78,10 +75,17 @@ def _map_click_to_segment(
         if patch.size == 0:
             return None
 
-        values, counts = np.unique(patch, return_counts=True)
+        # Native dai.SegmentationMask uses 255 as background/unassigned.
+        # Do not select the background as an object, while keeping class/instance
+        # id 0 valid.
+        foreground_patch = patch[patch != 255]
+        if foreground_patch.size == 0:
+            return None
+
+        values, counts = np.unique(foreground_patch, return_counts=True)
         return int(values[np.argmax(counts)])
 
-    def _send_mask(self, segmentation: SegmentationMask, mask: np.ndarray):
+    def _send_mask(self, segmentation: dai.SegmentationMask, mask: np.ndarray):
         mask_u8 = mask.astype(np.uint8) * 255
 
         out = dai.ImgFrame()

neural-networks/generic-example/AGENTS.md

@@ -86,7 +86,7 @@ Reusable single-model inference scaffold. It runs one Model Zoo model with one i
 ## Validation
 
 - `Run:` `python3 main.py`
-- `Alternative run:` `python3 main.py --model luxonis/mediapipe-selfie-segmentation:256x144 --overlay_mode`
+- `Alternative run:` `python3 main.py --model luxonis/mediapipe-selfie-segmentation:256x144`
 - `Archive run:` `python3 main.py --model /path/to/custom-model.tar.xz`
 - `Success looks like:` Visualizer exposes `Video` and `Detections`, and the pipeline runs until `q` is pressed
 - `Common failure meaning:` model identifier unavailable for platform, private model auth missing, or selected model violates the single-input/single-output assumptions

neural-networks/generic-example/README.md

@@ -23,7 +23,7 @@ Here is a list of all available parameters:
 -api API_KEY, --api_key API_KEY
                       HubAI API key for private HubAI access. Can also use 'DEPTHAI_HUB_API_KEY' environment variable instead. (default: )
 -overlay OVERLAY_MODE, --overlay_mode
-                      If passed, overlays model output on the input image when the output is an array (e.g., depth maps, segmentation maps). Otherwise, displays outputs separately.
+                      If passed, overlays model output on the input image when the output is an array (e.g., depth maps). Otherwise, displays outputs separately.
 ```
 
 ## Peripheral Mode
@@ -53,20 +53,26 @@ This will run a simple YOLOv6 object detection model (`luxonis/yolov6-nano:r2-co
 
 ```bash
 python3 main.py \
-    --model luxonis/mediapipe-selfie-segmentation:256x144 \
-    --overlay_mode
+    --model luxonis/mediapipe-selfie-segmentation:256x144
 ```
 
 This will run a selfie segmentation model.
 
 ```bash
 python3 main.py \
-    --model luxonis/yolov8-instance-segmentation-nano:coco-512x288 \
-    --overlay_mode
+    --model luxonis/yolov8-instance-segmentation-nano:coco-512x288
 ```
 
 And this will run an instance segmentation model.
 
+```bash
+python3 main.py \
+    --model luxonis/midas-v2-1:small-512x288 \
+    --overlay_mode
+```
+
+This will run a MiDaS depth estimation model and overlay the depth map on the input image.
+
 ```bash
 python3 main.py \
     --model /path/to/custom-model.tar.xz

neural-networks/generic-example/requirements.txt

@@ -1,5 +1,5 @@
-depthai==3.7.1
-depthai-nodes==0.5.0
+depthai==3.8.0
+depthai-nodes==0.6.0
 opencv-python-headless~=4.10.0
 numpy>=1.22
 python-dotenv

neural-networks/generic-example/utils/arguments.py

@@ -60,7 +60,7 @@ def initialize_argparser():
     parser.add_argument(
         "-overlay",
         "--overlay_mode",
-        help="If passed, overlays model output on the input image when the output is an array (e.g., depth maps, segmentation maps). Otherwise, displays outputs separately.",
+        help="If passed, overlays model output on the input image when the output is an array (e.g., depth maps). Otherwise, displays outputs separately.",
         required=False,
         action="store_true",
     )

neural-networks/object-detection/yolo-p/AGENTS.md

@@ -27,25 +27,24 @@ This is the repository reference for ADAS-style YOLO-P output: detections plus r
 - `Runs on:` RVC2 peripheral, RVC4 peripheral, and RVC4 standalone packaging
 - `Requires:` YOLO-P model and input resized to model dimensions
 - `Input:` camera frames by default or `ReplayVideo` via `--media_path`
-- `Output:` `Road Segmentation` and `Detections`
+- `Output:` `Video`, `Detections`, `Road Segmentation`, and `Lane Segmentation`
 - `Models:` YOLO-P YAMLs in [depthai_models/](depthai_models/)
 - `Visualizer / UI:` DepthAI Visualizer via `dai.RemoteConnection`
 
 ## Read First
 
 - [README.md](README.md)
 - [main.py](main.py)
-- [utils/annotation_node.py](utils/annotation_node.py)
 - [utils/arguments.py](utils/arguments.py)
 
 ## Architecture
 
 - `ParsingNeuralNetwork` runs the YOLO-P multi-head model on camera or replay input.
-- [utils/annotation_node.py](utils/annotation_node.py) consumes three outputs:
+- [main.py](main.py) sends the model outputs to Visualizer topics:
   - detections
   - road segmentation
   - lane segmentation
-- The Visualizer exposes the segmentation composite and detection overlay separately.
+- The Visualizer renders the input video, detections, road segmentation, and lane segmentation topics.
 
 ## Constraints
 
@@ -62,5 +61,5 @@ This is the repository reference for ADAS-style YOLO-P output: detections plus r
 ## Validation
 
 - `Run:` `python3 main.py`
-- `Success looks like:` the Visualizer shows `Road Segmentation` and `Detections`, with lane/road overlays and detected vehicles
+- `Success looks like:` the Visualizer shows `Video`, `Detections`, `Road Segmentation`, and `Lane Segmentation`, with road/lane segmentation and detected vehicles
 - `Common failure meaning:` the model output ordering changed, the replay input was resized incorrectly, or the operator expected a generic single-output detector

neural-networks/object-detection/yolo-p/main.py

@@ -4,7 +4,6 @@
 from depthai_nodes.node import ParsingNeuralNetwork
 
 from utils.arguments import initialize_argparser
-from utils.annotation_node import AnnotationNode
 
 _, args = initialize_argparser()
 
@@ -54,19 +53,11 @@
         input_node_out, nn_archive
     )
 
-    # annotation
-    annotation_node = pipeline.create(AnnotationNode).build(
-        frame=input_node_out,
-        detections=nn.getOutput(0),
-        road_segmentations=nn.getOutput(1),
-        lane_segmentations=nn.getOutput(2),
-    )
-
     # visualization
-    visualizer.addTopic(
-        "Road Segmentation", annotation_node.out_segmentations, "images"
-    )
-    visualizer.addTopic("Detections", annotation_node.out_detections, "images")
+    visualizer.addTopic("Video", nn.passthrough, "images")
+    visualizer.addTopic("Detections", nn.getOutput(0), "images")
+    visualizer.addTopic("Road Segmentation", nn.getOutput(1), "images")
+    visualizer.addTopic("Lane Segmentation", nn.getOutput(2), "images")
 
     print("Pipeline created.")
 

neural-networks/object-detection/yolo-p/requirements.txt

@@ -1,4 +1,4 @@
-depthai==3.7.1
-depthai-nodes==0.5.0
+depthai==3.8.0
+depthai-nodes==0.6.0
 opencv-python-headless~=4.10.0
 numpy>=1.22